Intent for Autonomous Research and Discovery Platforms

Domain Context: From Falkor to A-Lab

The autonomous research platform is no longer a thought experiment. The Schmidt Ocean Institute's R/V Falkor (too) operates ROV SuBastian and a growing suite of autonomous underwater vehicles on multi-week expeditions where science teams ashore direct sampling campaigns through high-latency satellite links while ship's officers retain authority over vessel safety and the chief scientist arbitrates instrument allocation. Lawrence Berkeley National Laboratory's A-Lab has demonstrated end-to-end autonomous synthesis and characterization of inorganic materials, with robots executing dozens of candidate compositions per day under high-throughput experimental designs proposed by machine-learning models. Self-driving laboratories at Berkeley, MIT, the University of Toronto, Carnegie Mellon, and commercial platforms such as Emerald Cloud Lab and Strateos extend the same pattern into chemistry, biology, and pharmaceutical discovery. Arctic and Antarctic research stations now operate instruments through entire dark seasons with only a skeleton crew physically present, with science teams in distant time zones issuing experimental directives that must be reconciled against station-keeper safety authority and host-nation regulatory constraint.

Across these settings, the same structural pattern recurs. Multiple authorities — each legitimate, each accountable to a distinct constituency — issue overlapping directives over the same physical apparatus. The principal investigator wants the experimental campaign completed. The lab manager wants the apparatus preserved and the safety case respected. The IRB or biosafety committee wants the scope of work to remain within the approved protocol envelope. The regulatory authority wants demonstrable evidence that none of the foregoing has crossed into territory requiring fresh regulatory review. Today these authorities are reconciled through a combination of email, calendar holds, lab-notebook entries, and bespoke if-then logic embedded in scheduling scripts. The reconciliation is implementationally-resolved and structurally-fragile.

The pattern intensifies as funding agencies push autonomous research toward higher consequence. The DOE Office of Science autonomous-laboratory roadmap, the DARPA Accelerated Molecular Discovery program, the NIH Bridge2AI cohort, and the European Open Research Cloud's autonomous-experiment workstreams all envision laboratory floors in which weeks of human-supervised work are compressed into days of unattended operation. ARIA in the United Kingdom and ARPA-H in the United States are funding programs whose deliverable is, in effect, a credentialed autonomous-discovery substrate; the sponsor is not buying experiments but the institutional capacity to run them safely without continuous human supervision. In the pharmaceutical sector, large discovery operations at Recursion, Insitro, Isomorphic Labs, and the AbCellera-class antibody-discovery platforms operate self-driving wet labs whose decisions cross into FDA-relevant territory at the moment a candidate moves toward IND-enabling studies. The authority-composition problem is no longer a research-administration nicety; it is a precondition for autonomous platforms to scale into regulated work.

Architectural Requirement

A research platform that operates autonomously across multi-day or multi-week campaigns must express, at the architecture layer, the answer to a set of questions that procedural compliance documents cannot answer in real time. Whose intent is currently authoritative for this instrument, at this moment, for this class of operation? When two authorities issue overlapping directives, by what rule is composition resolved? When a directive arrives from an authority whose credential has lapsed, expired, or been revoked mid-campaign, how does the platform refuse the directive without producing a hazardous abort? When an emerging-scenario sensor reading suggests the campaign envelope is being exceeded, which authority must be consulted before continuation, and within what latency bound?

These are not policy questions. They are architectural questions about how authority is represented, evaluated, and composed inside the control loop. A platform whose authority model exists only in human-readable protocols cannot answer them at machine timescales. The architecture must additionally express the temporal dimension of authority: an IRB protocol approved for twelve months expires; a PI's grant funding lapses on a fixed date and the directives written under it lose force at that moment; a lab manager's delegation to a graduate student is bounded by the student's training certification, which itself expires. Each of these temporal boundaries is, in current systems, a calendar reminder addressed to a human; in an autonomous platform it must be a structural admissibility condition that prevents the apparatus from acting on a directive whose authorizing credential has lapsed, even if the directive was queued days earlier and the credential lapsed mid-queue.

A further architectural requirement is reversibility-aware composition. The same campaign may contain operations whose consequences are fully reversible (re-run a measurement on the same sample), partially reversible (consume a small aliquot of a regenerable reagent), or irreversible (destroy a unique biological sample, release an organism into a containment-rated environment, publish a result to an external registry). The architecture must compose authority differently across these reversibility classes — a strategic intent from the PI may be sufficient for a reversible step, while an irreversible step may require simultaneous admissibility from PI, lab manager, and the responsible oversight committee. This composition is not expressible in a flat permission table; it requires the graduated-fidelity intent model the primitive supplies.

Why Procedural Compliance Fails

The procedural compliance approach — IRB protocol PDFs, standard operating procedures, signed change-control forms — was engineered for an era in which a human technician stood between every directive and every actuation. The technician supplied the authority-composition layer implicitly. They knew which PI's directive currently applied, which experiments lay outside the IRB scope, which instruments the lab manager had taken offline for maintenance, and which procedural guardrail required a hold for biosafety review. Their judgment was not formalized; it did not need to be.

Autonomous platforms remove the technician from that loop. Procedural compliance documents continue to exist, but they no longer participate in the moment-to-moment admission decision. What participates instead is whatever logic the platform's developers happened to encode — typically a flat priority hierarchy, a fixed scheduler, or a permissions table that bears no formal relationship to the institutional authority structure it is meant to represent. Drift between the procedural authority structure (what the IRB approved) and the implemented authority structure (what the scheduler actually enforces) becomes inevitable, undetectable in routine operation, and consequential only at the moment of an incident — at which point the audit trail consists of source code commits rather than authority records.

The drift is not malicious; it is the cumulative effect of ordinary engineering. A scheduler is patched to handle an edge case in a robotic-arm motion plan, and the patch incidentally relaxes a constraint that the original IRB protocol had assumed; six months later the platform is operating outside the protocol envelope and no one has noticed because the scheduler still passes its unit tests. A graduate student's training-certification expires; the credentials stay valid in the platform's local cache because no one wired the cache to the institutional credential authority, and the student's work continues under a credential that exists only inside the platform. A vendor firmware update changes the meaning of a calibration parameter, and a directive that previously stayed within safe operating bounds now exceeds them, but the procedural document references a parameter name not a parameter meaning. Each of these drifts is invisible to procedural review and would have been caught instantly by an architecture in which the institutional authority structure was the structural admissibility bound.

Investigations of recent autonomous-system incidents — across self-driving cars, hospital-pharmacy automation, and industrial robotics — share a common finding: the implemented authority structure had drifted from the procedural authority structure, and no one observed the drift until the incident. Autonomous research platforms are on the same trajectory, and the procedural compliance posture is no answer to it.

What Operator-Intent Provides

The operator-intent primitive supplies three structural elements that procedural compliance cannot. First, graduated fidelity tiers: intent is declared at a fidelity matched to the authority issuing it. A PI's high-level campaign goal is expressed as low-fidelity strategic intent; a lab manager's instrument-reservation envelope is expressed at mid-fidelity operational intent; a biosafety constraint is expressed as a hard admissibility bound. Each tier composes downward without requiring the higher tiers to enumerate every operational detail.

Second, multi-fleet and multi-authority intent fusion. A research platform that operates across a fleet of instruments — a robotic chemistry station, a characterization queue, a remote ROV, a station-keeping autonomous surface vessel — expresses intent fusion across the fleet such that an instrument-level admissibility decision considers not only the local directive stack but the institutional authority structure as a whole. A directive from an authority whose credential has lapsed produces a structural refusal, not a runtime exception.

Third, intent provenance that survives the campaign. Each admitted action carries forward the authority chain that admitted it: the PI directive, the IRB scope under which the directive was admitted, the lab-manager envelope active at the time, the regulatory perimeter in force. Post-hoc audit reconstructs the authority composition that produced each action without reverse-engineering it from logs.

The operator-intent primitive is the autonomous-platform specialization of the broader admission and intent substrate disclosed under USPTO provisional 64/049,409. The same five-property chain — credentialed observation, evidential weighting, composite admissibility, governed actuation with reversibility evaluation, and lineage-recorded provenance — applies to laboratory directives in the same way it applies to identity mutations or network-timing fixes. The graduation of fidelity (strategic / operational / hard-bound) is the property that lets a single substrate carry a PI's campaign aim and the IRB's species-and-protocol bound without forcing either to express itself at the granularity of a robot motion plan. The composition rule is structural, not stipulative: a candidate actuation is admitted if and only if the authority chain that admits it satisfies, at the relevant fidelity tier, every active intent in force at the moment of admission. The actuator surface itself is reversibility-aware, so the same substrate can run a reversible measurement under loose strategic admission while requiring a complete authority stack for an irreversible release.

Compliance Mapping

The graduated-fidelity intent declaration maps to IRB protocol scope: the approved protocol becomes a structural admissibility bound rather than a document referenced only at protocol-amendment time. Multi-authority intent fusion maps to the standard institutional separation between scientific authority (PI), operational authority (lab manager or chief scientist), oversight authority (IRB, biosafety committee, animal care committee), and regulatory authority (FDA for clinical-applicable work, EPA for environmental release, NOAA or flag-state maritime authority for ocean operations). Intent provenance maps to the audit-trail requirements that already apply to GxP-regulated research and that are emerging for federally funded autonomous-research programs. Where procedural compliance requires reconstruction, primitive-level compliance produces evidence as a byproduct of normal operation.

The compliance map extends across the regulatory landscape autonomous research is entering. 21 CFR Part 11 electronic-records requirements for FDA-regulated work are satisfied by the lineage record without bolting on a separate audit logger. The OECD Principles of Good Laboratory Practice that govern non-clinical safety studies are satisfied because each admitted action carries forward the authorizing study director's credential. NIH and DOE data-management-plan requirements are satisfied by the substrate's structural production of provenance metadata. The EU AI Act's high-risk-AI obligations, which begin to apply to AI-directed autonomous systems operating in regulated domains, are satisfied because the substrate itself produces the human-oversight and traceability evidence the regulation demands. The NIST AI Risk Management Framework's measurement and management functions map to the same lineage. In each case the procedural compliance posture requires building a separate evidence layer; the primitive produces the evidence as the byproduct of admission.

Adoption Pathway

Adoption proceeds in three stages. In the first, the operator-intent substrate is introduced as a parallel record alongside an existing scheduler — every admitted action is annotated with the authority chain that admitted it, without yet gating execution on the substrate. This produces an immediate audit-quality improvement and exposes drift between procedural and implemented authority. In the second, the substrate gates a defined subset of high-consequence operations: protocol-boundary actions, regulated-material handling, irreversible commits such as sample destruction or external release. In the third, the substrate becomes the platform's authoritative admission layer, with the legacy scheduler reduced to an execution engine that consumes admitted intent.

The first adoption surface is the platform-vendor layer: Emerald Cloud Lab, Strateos, Artificial, Synthace, Opentrons-class lab-automation vendors, and the lab-information-management-system incumbents (LabWare, STARLIMS, Benchling) all face the question of how their platforms will represent institutional authority once their customers move beyond single-PI research into multi-authority autonomous campaigns. The substrate is a feature their customers will demand once one competitor demonstrates audit-grade lineage as a built-in capability. The second adoption surface is the instrument-vendor layer: Thermo Fisher, Agilent, Waters, Bruker, and the robotic-chemistry vendors (HighRes Biosolutions, Hudson Robotics) who supply the actuators on autonomous floors will be asked to expose admission-aware interfaces so their hardware can participate in the chain rather than executing whatever directive arrives at the device API. The third adoption surface is the institutional layer: research universities, national laboratories, and pharma research operations whose IRBs, IACUCs, and biosafety committees increasingly recognize that paper protocols do not survive into autonomous regimes.

The convergence point is visible across pharmaceutical self-driving labs, materials-discovery platforms, ocean and polar research vessels, and the emerging class of AI-directed scientific-discovery systems. Each independently faces the authority-composition problem; each independently is rediscovering that procedural compliance does not extend into the autonomous regime. The operator-intent primitive positions a structural answer at that convergence point. Institutions that adopt the substrate before the first high-profile autonomous-research incident — the analog of the autonomous-vehicle fatalities, the hospital-pharmacy mis-dosing events, the industrial-robot injuries — acquire a defensible posture that institutions caught in the post-incident regulatory tightening will spend years trying to retrofit.